CN105718912A

CN105718912A - Vehicle characteristic object detection method based on deep learning

Info

Publication number: CN105718912A
Application number: CN201610052980.9A
Authority: CN
Inventors: 尚凌辉; 高勇; 刘小扬; 丁连涛; 李军伟
Original assignee: ZHEJIANG ICARE VISION TECHNOLOGY Co Ltd
Current assignee: ZHEJIANG ICARE VISION TECHNOLOGY Co Ltd
Priority date: 2016-01-26
Filing date: 2016-01-26
Publication date: 2016-06-29
Anticipated expiration: 2036-01-26
Also published as: CN105718912B

Abstract

The invention relates to a vehicle characteristic object detection method based on deep learning. All objects which need detecting in a vehicle window are marked, and a CNN (Convolutional Neural Network) structure is designed; a vehicle face is located roughly in a statistical machine learning method, and four points of the vehicle window are positioned to position the vehicle window; a suggestion frame which may be the suggestion frame of a target object is obtained in a selective search and edgebox combined suggestion frame extraction method; to-be-classified areas under different scales are obtained by carrying out multi-dimension zooming on the basis of the suggestion frame, and a multi-classification model is used for classification; prior information of each object itself is used to remove false detection; and the classes and coordinates of different characteristic objects of the windows of different vehicles are obtained. Compared with a traditional image processing method, the vehicle characteristic object detection method is higher in robustness, can detect all the objects of interest in the vehicle window in one step, and the speed is higher than that of a traditional learning method.

Description

A kind of vehicle characteristics object detecting method based on degree of depth study

Technical field

The invention belongs to technical field of machine vision, relate to a kind of vehicle characteristics object detecting method based on degree of depth study.

Background technology

By 2015, China's automobile pollution alreadyd more than 1.6 hundred million.Along with expanding economy, this numeral is also in sustainable growth.Substantial amounts of automobile travels on road, to vehicle supervision department with huge government pressure., in many case involving public securities, there is the situation that automobile occurs as the vehicles in the opposing party.

Automobile management automatization means existing are mainly electronic police and bayonet system by our country.These systems can captured in real time vehicle high definition picture, and automatically analyze out the information such as the number-plate number, vehicle information (include brand, model and time, currently progressively implement), color, reach the automated management to vehicle and quickly search.But namely allow to make full use of aforementioned information, it is also difficult to found out by specific suspected vehicles, especially when the picture not having suspected vehicles, and when only having eye witness.Hence with on car except above-mentioned information, the exclusive structural description information of each car, whether put down including sunshading board, paper towel box, suspension member, identifier etc., become the important clue determining particular vehicle.

Current prior art focuses primarily upon the detection of vehicle sun visor, including traditional image processing method, such as " a kind of vehicle sun visor detection method based on graphical analysis and device-201210089548.9 ", " detection method-201310365024.2 of a kind of vehicle sun visor state " and " a kind of car internal sunshade board detecting method and device-201310574043.6 ", with conventional machines learning method, such as " a kind of vehicle sun visor condition detection method and device-201510531752.5 ", " vehicle sun visor detection method and device-201310512222.7 based on graphical analysis ".Wherein, adopting traditional image processing method for environmental change, such as illumination variation, when reflective, effectiveness comparison is poor, it is easy to cause flase drop.And traditional machine learning method needs sunshading board feature is manually extracted.Owing to sunshading board feature is relatively single, therefore, cause that traditional machine learning method is excessively poor to the Detection results of some extreme environments.On the other hand, conventional machines learning method utilizes the method for scan box that all possible region of different scale hypograph is scanned, speed very slow (often several ten thousand scan box of a figure).One object all can only be detected by existing method.

Summary of the invention

In order to overcome the traditional method can only to certain object detection in vehicle window, the present invention adopts the method that Suggestion box extractive technique learns to combine with the degree of depth, objects multiple in vehicle window are detected simultaneously, thus reaching to detect object as much as possible within the shortest time, to meet the requirement of real-time that data are processed.

The inventive method comprises the following steps:

Step 1. all marks out the object that be there is a need to detection in vehicle window, design CNN network structure, this network mainly includes three convolutional layers, three pond layer (1 Max pond layers, 2 AVE pond layers), three RELU layers, two full articulamentums and Softmaxwithloss layer, and be trained on Caffe framework by this structure, obtain disaggregated model more than.

Step 2. is by the position of the method coarse localization car face of statistical machine learning, and four points of vehicle window are positioned, and obtains the position of vehicle window.

Step 3., in conjunction with selectivesearch (SelectiveSearchforObjectRecognition) and edgebox (EdgeBoxes:LocatingObjectProposalsfromEdges) Suggestion box extracting method, obtains being probably the Suggestion box of target object；This Suggestion box extracting method mainly generates the initialization block of selectivesearch first with edgebox, owing to edgebox own splits image with image border, therefore, have based on the initialization area of edgebox more accurate than the initialization area method in former selectivesearch method.

Step 4. carries out the region to be sorted that multiple dimensioned convergent-divergent obtains under different scale on the basis of Suggestion box, then utilizes many disaggregated models that step 1 obtains to classify.

The prior information that step 5. utilizes each object own carries out the last flase drop that goes and processes.Here utilizable prior information includes the position probability distribution figure of each object, size, length-width ratio etc..According to these prior informations, it is possible to effectively remove some flase drops so that it is more accurate to detect.

Step 6. obtains the various characteristic body classifications on each car vehicle window and coordinate.

Beneficial effects of the present invention: the relatively conventional image processing method of the present invention has higher robustness, and disposable can detect all objects interested in vehicle window, and speed (Suggestion box every Figure 200 about 0 Suggestion box faster than conventional learning algorithms, several ten thousand even tens0000 Suggestion box are often schemed relative to tradition scan box), it is possible to automatically provide the feature with distinction for vehicle retrieval system and erect word and describe and the bridge of picture conversion.

Accompanying drawing explanation

Fig. 1 is the inventive method flow chart.

Fig. 2 is SelectiveSearch(SS) flow chart；

Fig. 3 is RCNN flow chart.

Detailed description of the invention

Below in conjunction with embodiment and accompanying drawing, the invention will be further described:

Overall process of the present invention is specifically shown in Fig. 1, is described in detail as follows:

1., owing to the present invention needs training in advance one based on many disaggregated models of CNN, training flow process is shown in the right half part of Fig. 3, and training detailed process is described below:

1) marking the data collected, currently mainly mark sunshading board, annual inspection mark accords with, suspension member, goods of furniture for display rather than for use, paper towel box five class.

2) take all of positive sample and zoom to unified size, such as 32x32, calculate all sample averages, then all samples deduct average.

3) design CNN network structure.The present invention mainly adopts following network structure:

31) convolutional layer

32) Max pond layer

33) Relu layer

34) convolutional layer

35) AVE pond layer

36) PRelu layer

37) convolutional layer

38) AVE pond layer

39) PRelu layer

310) full articulamentum

311) full articulamentum

312) Softmaxwithloss layer

Wherein, softmaxwithloss uses softmax function computing formula as follows:

,

Represent classification, can be determined that certain object z belongs to the probability size of the i-th class by this.

4). the many disaggregated models of CNN of vehicle characteristics detection are obtained by Caffe training.

2. collect an image by monitor video.

3. the positional information on four summits of vehicle window it is accurately obtained by statistical learning method and positioning feature point.

4. likely comprised the Suggestion box of object in conjunction with selectivesearch and edgebox, the method flow process is specifically shown in Fig. 2,

1) utilize edgebox to calculate the edges of image, then obtain an edgegroups, obtain m prime area according to these edgegroup, be designated as.One similarity set is set and is designated as S, be initialized as sky.

2) similarity in each two region is calculated, and be deposited in S set.

3) two regions that similarity is maximum are found out,, delete the similarity in these two pieces of regions and other regions, be then combined with this two region, and recalculate this region and adjacent similarity, be deposited into S；Repeat this process until the number in S collection is less than number predetermined in advance always.

4) the 3rd is extracted) region that obtains of step is as the Suggestion box being likely target.

5. obtaining different Suggestion box by the 4th step in conjunction with multiple dimensioned, utilize the many disaggregated models of CNN that the 1st step obtains that different Suggestion box is calculated, it was predicted that the classification of each Suggestion box and belong to the confidence level of the category, this flow process refers to the left-half of Fig. 3.

6. the information according to statistics, in conjunction with confidence level, removes some flase drops.Utilizable information includes the Position probability densities distribution of every type objects, size, length-width ratio etc..

7. obtain the object and the position coordinates that contain on last each vehicle window.

The above; it is only presently preferred embodiments of the present invention, is not intended to limit protection scope of the present invention, should understand by band; the present invention is not limited to implementation as described herein, and the purpose that these implementations describe is in that to help those of skill in the art to put into practice the present invention.

Claims

1. the vehicle characteristics object detecting method based on degree of depth study, it is characterised in that the method comprises the following steps:

Step 1. marks out the object that be there is a need to detection in vehicle window, design CNN network structure, this network structure mainly includes three convolutional layers, three pond layers, three RELU layers, two full articulamentums and Softmaxwithloss layer, and be trained on Caffe framework by this structure, obtain disaggregated model more than；

Step 2. is by the position of the method coarse localization car face of statistical machine learning, and four points of vehicle window are positioned, and obtains the position of vehicle window；

Step 3., in conjunction with selectivesearch and edgebox Suggestion box extracting method, obtains being probably the Suggestion box of target object；

Step 4. carries out the region to be sorted that multiple dimensioned convergent-divergent obtains under different scale on the basis of Suggestion box, then utilizes many disaggregated models that step 1 obtains to classify；

The prior information that step 5. utilizes each object own carries out the last flase drop that goes and processes；

2. according to claim 1 a kind of based on the degree of depth study vehicle characteristics object detecting method, it is characterised in that: step 3 specifically:

1) utilize edgebox to calculate the edges of image, then obtain an edgegroups, obtain m prime area according to these edgegroup, be designated as；One similarity set is set and is designated as S, be initialized as sky；

2) similarity in each two region is calculated, and be deposited in set S；

3) two regions that similarity is maximum are found out,, delete the similarity in these two pieces of regions and other regions, be then combined with this two region, and recalculate this region and adjacent similarity, be deposited into set S；Repeat this process until the number in S collection is less than number predetermined in advance always；

3. a kind of vehicle characteristics object detecting method based on degree of depth study according to claim 1 and 2, it is characterised in that: described prior information includes the position probability distribution figure of each object, size, length-width ratio.